Intelligence is an API now

A marketer reviewing agency copy against brand guidelines is doing it. A broker chasing documents and summarising a case is doing it. An ops manager turning last week's numbers into a Monday summary is doing it. The nouns change; the shape doesn't.

For the entire history of office work, the middle step was the expensive part. Judgement lived only in people, so every one of these loops needed a person in it, and the loops queued up behind the people.

That's what changed. Judgement, of the decent, fast and tireless kind, is now available on demand, the way electricity or web hosting is. The industry calls this "intelligence as an API". In practice it means you can call reading-and-deciding as a service, thousands of times a day, for pennies.

So why doesn't it feel that way at work?

Because of the first and third steps.

Models are genuinely good at the middle one. But out of the box they arrive knowing nothing about your business and connected to nothing in it. The hard part was never "can the machine think?". It's getting your information to the thinking, and getting the output back into the work: into the CRM, the reply, the report, with someone approving it before it counts.

That's plumbing. Unglamorous, very buildable plumbing: connectors into the tools you already use, working environments where the AI can see the right documents, and human approval at the moments that matter. Teams that feel "AI isn't there yet" are almost always missing plumbing, not intelligence.

The other missing piece: your way of doing things

There's a second gap, subtler than plumbing. Generic intelligence gives generic output.

The fix isn't to re-explain yourself in every chat. Modern AI setups let you write things down once and have them apply every time: your brand voice, your definitions, your process for handling a complaint, the format your boss actually reads. Different tools call these skills, instructions, or memory. The mechanics matter less than the idea: the useful version of AI isn't the model; it's the model plus an accumulating pile of your own know-how, written down where it can act on it.

What I'd actually do with this

Not a grand programme. Three moves, small enough to start this week.

Find one loop. One task with the gather → judge → produce shape that happens weekly or daily. The boring, reliable ones are the best candidates: a report, a triage, a first draft of a reply.

Wire it, don't retype it. Connect the AI to where that information already lives instead of copy-pasting fragments into a chat window. The difference in usefulness is the difference between a tourist and a colleague.

Write down the judgement. Capture how a good version of that task is done: what to check, what tone, what counts as an exception a human must see. That document is an asset that works every time the loop runs.

Then measure whether it helped, and only then pick the next loop.

That is, more or less, the whole discipline of putting AI to work: one loop at a time, wired properly, taught your way of doing things, with a person approving what leaves the building. The intelligence is already there, waiting behind an API. The advantage now goes to whoever does the wiring.

That wiring is what I do, for my own work every day and for teams that want it done to theirs.